X Algorithm Update: The New-Post Cold-Start Window Is Now Two Hours

A set of default parameters synced to X's official open-source recommendation repository in late September 2026 materially tightened the new-post cold-start mechanism. The author follower cap moved from 1,000 to 50,000; maximum post age fell from 48 hours to two; the Home impression threshold dropped from 1,000 to 200; and Thompson sampling was enabled. But 'the first two hours decide life or death' is an operating reminder, not a literal system rule. After two hours, a post loses eligibility for this specific cold-start lift; it may still be discovered through follows, Phoenix, Thunder, SimClusters, reposts, search or later engagement.

• Cold-start retrieval cap: PhoenixColdStartMaxResults changed from 0 to 200
• Home impression threshold: ColdStartImpressionThreshold fell from 1,000 to 200
• Author follower cap: ColdStartFollowerCap increased from 1,000 to 50,000
• Maximum post age: ColdStartMaxPostAgeSecs fell from 172,800 seconds to 7,200 seconds—from 48 hours to two
• Lift target: default slot parameters are 15 and 16; the code samples from a half-open range, so it is inaccurate to promise a fixed 'position 16' every time
• Exploration method: EnableColdStartThompsonSampling was turned on; the default samples a top two before using the underlying ranking score to select the post that receives lift

The public parameter caps Phoenix cold-start retrieval at 200 candidates, but that is a maximum—not a guarantee that every request inserts 200 new posts. What reaches ranking and receives lift still depends on retrieval, filtering, experiment assignment, eligibility and the request itself. The draft's claims of 'fewer than eight likes' and 'fewer than 500 views' do not appear in the current eligibility code. Confirmed conditions include: the post is neither a reply nor repost; the author has no more than 50,000 followers; the post is no older than two hours; Home impressions are below 200; and corpus and position conditions are satisfied. Unsupported thresholds should not be presented as platform rules.

The cold-start selector feeds fav_count and view_count_on_home into a Beta distribution. In practical terms, the system balances exploitation of observed performance with exploration randomness, so a post with little data is not automatically excluded. Uncertainty is higher with small samples; as evidence accumulates, observed favourite behaviour matters more. The selected candidate is still resolved with the underlying model score. More importantly, the full For You ranker does not use likes alone. X's public architecture describes predicted replies, reposts, shares, dwell, follows and negative feedback, among other actions. Operators should not turn every post into a request for likes.

Raising the follower cap from 1,000 to 50,000 means cold-start lift is no longer reserved for very small accounts. A new account may receive a test, but it may also compete with small and mid-sized creators that have stronger content skills, audience context and historical signals. A low follower count is not an advantage by itself, and a larger audience is no guarantee. The practical question is whether the first relevant viewers can quickly understand the post and respond authentically.

When cold-start eligibility is limited to two hours, whether the initial audience is online can affect the quality of early feedback. But there is no universal timing chart. A B2B account can test commute, lunch and after-work windows in the target market; a consumer account should segment by geography, interest and use case. Record at least four weeks of data and separate time zone, weekday, format and topic. Otherwise, a random spike can easily be mistaken for an algorithm rule.

• State the problem or conclusion in the opening so the post is understandable without expanded context
• Give one post one job; remove unrelated claims and links
• Use specific data, experience, comparisons or evidence to reduce promotional vagueness
• Make the media and copy communicate the same idea; do not let the first-frame promise diverge from the body
• Prepare to answer real questions, but do not paste identical replies or manufacture activity with irrelevant comments
• Links are not inherently forbidden, but a post that requires leaving X to understand its value creates additional early-stage friction

• Confirm that rendering, media, links and line breaks work; fix technical errors promptly
• Answer genuine questions and add useful information so the conversation keeps creating value
• Record Home impressions, likes, replies, reposts, profile visits and follows at 15 minutes, 30 minutes, one hour and two hours
• Separate two-hour results from 24- and 48-hour performance to distinguish cold-start behaviour from long-tail distribution
• Do not coordinate likes, buy engagement, use secondary accounts to boost a post or mass-copy replies; X's authenticity policy prohibits coordinated engagement exchange and artificial amplification
• Do not immediately delete and repost a weak result; diagnose topic, opening, audience, timing and expression before the next test

• Content context: publish date, time zone, topic, format, length, media and outbound links
• Early outcome: Home impressions and engagement at 15 minutes, 30 minutes, one hour and two hours
• Quality outcome: substantive replies, profile visits, new follows, website visits and qualified enquiries
• Long-tail outcome: incremental impressions, search discovery, reposts and citations at 24 hours, 48 hours and seven days
• Test design: test several publish windows for the same topic, and different openings and structures within the same window
• Review rule: record only changes supported by evidence; do not turn one viral post into a universal law

The current public code includes interest-based candidate sources such as SimClusters and adds a SID similarity source that is disabled by default. This shows that retrieving related content from behavioural context is a clear system direction, but a disabled default must not be reported as fully launched. For operators, topic consistency makes it easier for people and systems to understand the problems an account repeatedly solves. It does not replace content quality, audience fit or authentic engagement.